The idea that Dave is talking about is that I don't think we should be doing any classloader special sauce in accumulo-start at all, and we might even be able to remove accumulo-start as a module entirely, since this is its primary (sole?) purpose.
It's just a rough idea that I've tossed around with a few people, but haven't really spent any time materializing it into a proposal, PR, or experiment. Basically, I think we should rip out all classloader special sauce. If a user still wishes to use a custom classloader for any reason, using vfs2 or anything else, they can set a system class loader with -Djava.system.class.loader=my.custom.CustomClassLoader when they run java. This is an advanced Java option supported by Java itself, and really shouldn't be a problem to punt this downstream. Classloading is way outside the scope of what Accumulo does anyway, and Accumulo should have its complexity centered around what it does, and not "bells and whistles" on top of basic Java language functions. If we wanted to, we could use our current classloading code to create a classloader which could be used this way... and maybe provide it as an example or explain it in a blog post. But, Accumulo shouldn't be doing special sauce class loading... there are other projects that are better suited to specializing that for any Java application... and there's no reason we need it so tightly coupled to Accumulo. Of course, there's still some utility in the per-table context classloaders for pluggable components like iterators... and there's probably room for improvement in the configuration of those... but the main startup classloading is probably best to rip out. I'm not sure if it should be done for 2.0 or not... maybe yes. I'd be willing to rip it out... I enjoy ripping things out and reducing code complexity. But, I don't really have a desire to do the work of implementing or blogging about alternatives, if that's even necessary. I'd hope that somebody else would do that, if they felt it was really necessary once the built-in stuff was ripped out. For me, I'd be happy mentioning the feature in the release notes, maybe linking to the docs on the feature, and leaving implementation as an exercise for downstream, with an open invitation for a guest blog on our website about how it could be done. I've been thinking we're probably going to want a second alpha... or a beta, before 2.0 final... and if we did this for 2.0, I'd definitely want another pre-release release first. On Wed, Oct 24, 2018 at 3:10 PM Dave Marion <[email protected]> wrote: > > I have talked with Christopher about the VFS class loader in general and I > think he has a good approach. He can elaborate further if needed, but the > approach is to move it out of the core project and allow users to configure > it at runtime using the java.system.class.loader system property. There are > organizations using the VFSClassloader successfully, maybe it just needs to > be reimplemented. > > On Wed, Oct 24, 2018 at 2:58 PM Sean Busbey <[email protected]> > wrote: > > > sounds like a good DISCUSS thread for 2.0? > > On Wed, Oct 24, 2018 at 1:43 PM Josh Elser <[email protected]> wrote: > > > > > > It seems like commons-vfs2 is just a pile of crap. > > > > > > It's been known to have bugs for years and we've seen zero progress from > > > them on making them better. > > > > > > IMO, rip the whole damn thing out. > > > > > > On 10/24/18 12:42 PM, Matthew Peterson wrote: > > > > Hello Accumulo, > > > > > > > > Summary: commons-vfs2 version 2.2 seems to have problems and it may be > > > > worth rolling back to version 2.1 of commons-vfs2. > > > > > > > > My project upgraded a system from Accumulo 1.8.1 to 1.9.2. Immediately > > > > after switching vfs contexts we saw problems. The tservers would > > error in > > > > iterators about missing classes that were clearly on the classpath. > > The > > > > problems were persistent until we replaced the commons-vfs2.jar with > > > > version 2.1 (Accumulo 1.9.2 uses version 2.2). Until we rolled vfs > > back, > > > > we received errors particularly with Spring code trying to access > > various > > > > classes and files within the jars. It looks like in 2.2, commons-vfs > > > > implemented a doDetach method which closed the zip files. We suspect > > that > > > > code is the problem but haven't tested that theory. > > > > > > > > I suspect that most users don't use this feature. > > > > > > > > Thanks! > > > > Matt > > > > > > > > > > > > -- > > busbey > >
